Segmentation, classification and modelization of textures by means of multiresolution decomposition techniques

Abstract

An interesting problem in computer vision is the analysis of texture images. In this work, we have developed specific methods to solve important aspects of this problem. The first approach involves segmentation of a specific type of textures, i.e. those of microscopy images of thin marble sections. These images comprise a pattern of grains whose sizes and shapes help specialists to identify the origin and quality of marble samples. To identify and analyze individual grains in these images represents a problem of image segmentation. In essence, this involves identifying boundary lines represented by valleys which separate flat areas corresponding to grains. Of several methods tested, we found those based on mathematical morphology particularly successful for segmentation of petrographical images. This involves a pre-filtering step for which again several approaches have been explored, including multiresolution algorithms based on wavelets. In the second approach we have also used multiresolution analyses to address the problem of classifying texture images. In contrast to more global approaches found in the literature, we have explored situations where visual differences between textures are rather subtle. Since we have tried to impose relatively few restrictions on these analyses, we have developed strategies that are applicable to a wide range of related texture images, such as images of ceramic tiles, microscopic images of effect pigments, etc. The approach we have used for the classification of texture images involves several technical steps. We have focused our attention in the initial low-level analyses required to identify the general features of the image, whereas the final classification of samples has been performed using generic classification methods. To address the early steps of image analysis, we have developed a strategy whereby the general features of the image fit one of several pre-defined models with increasing levels of complexity. These models are associated to specific algorithms, parameters and calculations for the analysis of the image, thus avoiding calculations that do not provide useful information. Finally, in a third approach we want to arrive to a description of textures in such a way that it should be able to classify and synthesize textures. To reach this goal we adopt a probabilistic model of the texture. This description of the texture allows us to compare textures through comparison of probabilistic models, and also use those probabilities to generate new similar images. In conclusion, we have developed strategies of segmentation and classification of textures that provide solutions to practical problems and are potentially applicable with minor modifications to a wide range of situations. Future research will explore (i) the possibility of adapting segmentation to the analysis of images that do not necessarily involve textures, e.g. localization of subjects in scenes, and (ii) classification of effect pigment images to help identify their components.

Date of Award	1 Oct 2001
Original language	Undefined/Unknown
Supervisor	Joan Serrat Gual (Director)